J-2013-NEUROCOM-Multi-Instance Multi-Label Image Classification A Neural Approach

《J-2013-NEUROCOM-Multi-Instance Multi-Label Image Classification A Neural Approach》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、Neurocomputing99(2013)298–306ContentslistsavailableatSciVerseScienceDirectNeurocomputingjournalhomepage:www.elsevier.com/locate/neucomMulti-instancemulti-labelimageclassification:Aneuralapproacha,naa,ba,cZenghaiChen,ZheruChi,HongFu,DaganFengaCentreforMultimediaSignalProcessing,DepartmentofElectronica

2、ndInformationEngineering,TheHongKongPolytechnicUniversity,HungHom,Kowloon,HongKongbDepartmentofComputerScience,ChuHaiCollegeofHigherEducation,TsuenWan,HongKongcSchoolofInformationTechnologies,TheUniversityofSydney,NSW2006,AustraliaarticleinfoabstractArticlehistory:Inthispaper,amulti-instancemulti-la

3、belalgorithmbasedonneuralnetworksisproposedforimageReceived22February2012classification.Theproposedalgorithm,termedmulti-instancemulti-labelneuralnetwork(MIMLNN),ReceivedinrevisedformconsistsoftwostagesofMultiLayerPerceptrons(MLP).Formulti-instancemulti-labelimage28July2012classification,alltheregiona

4、lfeaturesarefedtothefirst-stageMLP,withoneMLPcopyprocessingAccepted1August2012oneimageregion.Afterthat,theMLPinthesecondstageincorporatestheoutputsofthefirst-stageCommunicatedbyX.LiMLPstoproducethefinallabelsfortheinputimage.Thefirst-stageMLPisexpectedtomodeltheAvailableonline8August2012relationshipbetw

5、eenregionsandlabels,whilethesecond-stageMLPaimsatcapturingthelabelKeywords:correlationforclassificationrefinement.ErrorBack-Propagation(BP)approachisadoptedtotunetheMulti-instancemulti-labellearningparametersofMIMLNN.Inviewofthattraditionalgradientdescentalgorithmsuffersfromlong-termImageclassification

6、dependencyproblem,arefinedBPalgorithmnamedRpropisextendedtoeffectivelytrainMIMLNN.NeuralnetworksTheexperimentsareconductedonasyntheticdatasetandtheCoreldataset.ExperimentalresultsSyntheticdatademonstratethesuperiorperformanceofMIMLNNcomparingwithstate-of-the-artalgorithmsformulti-instancemulti-labeli

7、mageclassification.&2012ElsevierB.V.Allrightsreserved.1.Introductionmultiplelabels.ButMLLregardsanimageasasingleinstance.Theregionalfeaturesarenotemployed.TheissueisthattheWiththeprevalenceofdigitalima